Multi-column deep neural networks for image classification

The work targets image recognition tasks such as handwritten digit and traffic sign classification, where conventional computer vision and machine learning had not matched humans. Its architecture is a wide, deep, biologically plausible convolutional network built from winner-take-all neurons with small receptive fields, giving a depth comparable to the many layers between retina and visual cortex in mammals. Only winner neurons are trained, several deep columns are each specialized on inputs preprocessed in different ways, and their predictions are averaged, with graphics cards used to make training fast.

On the competitive MNIST handwriting benchmark the method was the first to reach near-human performance, and on a traffic sign recognition benchmark it surpassed humans by a factor of two, while also improving the state of the art across many common image classification benchmarks. These results showed that deep, GPU-trained convolutional ensembles could close the gap to human accuracy on hard visual recognition problems.